XAI slashes brain scan time from 27 to 14 minutes on Connectome 2.0

A team from multiple institutions leveraged an Explainable AI (XAI) framework to drastically shorten a specialized brain MRI protocol. The Neurite Exchange Imaging (NEXI) model, which probes gray matter microstructure by measuring compartment diffusivities, neurite fraction, and exchange time, typically requires prohibitively long multi-shell, multi-diffusion-time acquisitions. By training XGBoost with SHAP and Recursive Feature Elimination on synthetic signals, the researchers identified an optimal subset of just 8 MRI features—down from the original 15. This cut scan time from 27 minutes to 14 minutes on the powerful Connectome 2.0 scanner.

Validated against the full acquisition and theoretical benchmarks, the XAI protocol not only matched the Cramér-Rao Lower Bound optimum but also proved more robust than heuristic alternatives like "Mid-Range" (which biased exchange time estimates) and "Corner" (which showed 5-fold higher variability in intra-neurite diffusivity). This model-agnostic optimization approach is easily adaptable to numerical models or complex noise scenarios where CRLB becomes intractable. The 14-minute protocol accelerates exchange-sensitive microstructural mapping, making it viable for clinical translation on existing scanners.